on 
| vectors 
der 
19 fields. 
on 
id in 
  
at 
le, 
inel, 
the left part of the image, where both water classes are misclassified as forest. The 
direction dependant statistics yield a marked improvement without errors near the 
edges (figure 4 b) and produce a homogeneous classification over the entire strip. The 
accuracies of training fields increases for grass from 70 % to 99 96, for coniferous 
forest from 80 96 to 100 96, for lake from 44 96 to 83 96. A detailed analysis of 
differences between both direction dependant classifications show few differences. 
They can be explained by differences in the training area selection which was done by 
two persons. 
4. CONCLUCIONS ' 
The results show that a considerable improvement in the classification of airborne 
multispectral scanner data over the entire scan angle can be achieved using direction 
dependant statistics representative for the whole strip. 
The suggested method uses polynomial coefficientsto calculate these statistics. 
For homogeneous unvegetated objects (bare soil , water) about 20 - 30 uniformly 
distributed training fields are sufficient to determine the polynomials. If objects with 
similar spectral characteristics or vertical structure should be separated, about 60 
training fields are needed to determine also trendcorrected covariance matrices. Similar 
to non-direction dependant classification the accuracy of direction dependant 
classification. is considerable influenced by the reliability of training area selection. 
Therefore other methods to determine statistical features should be developed in the 
future. 
5. LITERATURE 
[1] PFEIFFER, B.: Untersuchung des richtungsabhüngigen Stählungevethältens in 
multispektralen Abtastdaten. Bildmessung und Luftbildwesen 
50 (1982), S. 35 - 47 
[2] SWAIN, P.H., DAVIS, S.M.: Remote Sensing - The Quantitative Approach 
Mc Graw-Hill, New York, 1978 
193 
dümo o s S S SS ORDRE RR EUER REED EE Act